Google Antigravity Guide: Install to Your First Agent Task

People who search for antigravity or anti gravity are usually not looking for just another chatbot. They want to know whether Google Antigravity can directly help change code, run tests, operate a browser, and where a beginner should start. Google defines Antigravity as an agent-first development platform. It provides a familiar editor experience along with the Manager Surface, which is used to dispatch and observe long-running tasks. This article is based on Google's official materials as verified on September 21, 2026. Downloads, sign-in eligibility, and the features actually shown in your account are still determined by the official pages.

1. What Is Google Antigravity?

Antigravity is more than a completion button inside an editor. Google's introduction describes it as an agentic development platform that lets agents plan, execute, and verify fairly complete development tasks across the editor, the terminal, and the browser.

This positioning differs from the usual flow of "ask a question and get a piece of code." You can assign a task with clear completion criteria, for example, "Read the current login flow first, list the risks, and then propose a minimal test plan." The agent organizes its work steps, may read files or run commands, and finally reports back with results you can inspect. On first use, it is best to start with read-only, analysis, or test tasks that will not change any data.

It is also not an automatic deployment tool that will always finish a whole project correctly once you hand it over. An agent may still misunderstand requirements, pick the wrong files, or produce changes that need human review, so the task boundaries and acceptance criteria should be written down clearly first.

2. How Do Editor View and Manager Surface Differ?

Google divides how you work with Antigravity into two main surfaces:

• Editor View: suited to synchronous work where you read code yourself, edit files, and use inline commands and real-time completion.
• Manager Surface: suited to dispatching, coordinating, and observing asynchronous work where multiple agents run in different workspaces.

You can think of them as the "workbench" and the "task control panel." Small fixes, line-by-line review, and quick experiments usually start in Editor View. When a longer repair, test, or maintenance task needs to be split off, consider the Manager Surface. The benefit is that you do not have to turn all work into background tasks, and it is easier to keep track of the scope of changes.

If you want to understand how AI development tools and the division of labor for API and model selection fit together, you can read the Unified AI Model Entry for Engineering Teams on this site. That article looks at architecture and integration, not at Antigravity's installation documentation.

If you are not yet familiar with the division of labor between "models, APIs, and applications," read the AI API Beginner's Guide first, then come back to decide which work should go to a development agent.

3. How Should Beginners Complete a First Task?

The official download entry point is antigravity.google/download. Which operating systems, accounts, and models the product supports may change with the preview release; this article does not fix the screens or regional eligibility of any particular download page.

After installation, you can start with the following process:

1. Open a restorable project workspace, and first confirm the current branch and any uncommitted changes.
2. Assign a task with a very small scope, for example, "Only read src/, find the current test entry point and list the run commands, and do not modify any files."
3. Ask the agent to report its plan first, including the paths it will read or execute, and the completion criteria.
4. After reviewing the plan, allow it to make one small, verifiable change; when finished, ask it to list the modified files, the test results, and any open issues.
5. Check the diff, the test output, and the actual screens yourself, and then decide whether to keep the changes.

This order may seem slow, but it suits getting familiar with an agent's working scope for the first time. Once you know how it understands the workspace, commands, and acceptance criteria, handing it tasks to run in the background will be safer.

4. How Should You Read Artifacts?

Google places particular emphasis on Artifacts, rather than only showing users the raw tool log. Artifacts can be task lists, implementation plans, screenshots, or browser recordings, letting you quickly judge from the deliverables whether the agent is heading in the right direction.

When reviewing an Artifact, you can ask four questions:

• Does it clearly list which files were actually modified?
• Do the screenshots or recordings really correspond to the requirements, rather than just showing that it "looks finished"?
• Do the tests cover the behavior you asked for, rather than only confirming that it starts successfully?
• What assumptions, errors, or unfinished items remain?

If the answers are incomplete, you can add feedback directly on the Artifact and ask the agent to fix the work or provide more evidence. For work that needs to be handed over, Artifacts are also better review material to keep than a long, unreadable command log.

5. What Should You Watch Out for When Using Antigravity?

If an agent can work across the editor, the terminal, and the browser, it may access more project content than a typical chat window. Before you start, it is recommended that you:

• Do not paste API keys, passwords, or personal data into prompts; use the project's existing secrets management.
• Test first with a minimal workspace and minimal permissions, and then gradually expand the scope it can read or execute.
• Set manual confirmation checkpoints for actions such as deletion, deployment, database writes, and payments.
• Do not rely only on the agent's summary; at minimum, check the diff, the tests, the browser results, and the actual output.
• If you use a preview feature, first confirm the conditions, regions, and available models listed on the current product page.

At the 2026 I/O developer keynote, Google mentioned directions such as a cross-platform terminal sandbox, credential masking, and hardened Git policies. This article has not verified step by step whether your account and the current version display each of these controls, so they are marked unverified. Do not treat the security recommendations in this article as meaning that all protections are already enabled automatically in your workspace.

6. Which Related Model IDs Does BazaarLink Have?

Before writing this article, we queried https://bazaarlink.ai/api/v1/models on September 21, 2026. The public list showed gemini-3.8-flash and gemini-3.1-pro-preview, but did not show antigravity-preview-05-2026. For that reason, this article does not claim that the model options inside the Antigravity product are models BazaarLink provides.

If you are building a tool outside Antigravity that needs the Gemini API, these IDs found in the public list are candidate values, and you should query the public list again before any actual request. This note only states which model IDs could be found. It does not mean Antigravity needs to connect to BazaarLink, and it does not change Antigravity's own sign-in or model settings.

Official Verification Sources (verified: September 21, 2026)

• Google Developers Blog: Build with Google Antigravity
• Google Developers Blog: Google I/O 2026 Developer Keynote
• Google AI for Developers: Gemini API models
• Google Antigravity official download page
• BazaarLink public model list

Unverified items: download and sign-in eligibility by region and account as of September 21, 2026; the model menu actually offered in the current version; whether all security controls are enabled in your workspace; and the compatibility between the Antigravity product and BazaarLink endpoints. This article does not cite prices.

TWD billing · Taiwan invoices · leading AI models · OpenAI-compatible API